Milan, AI and Questions That Won’t Let Me Go
I spent two days at AI Week in Milan, the biggest European AI event of the year. I came home with pages of notes, several key ideas, and one firm conviction: companies, not just those in my portfolio, must stop thinking about AI as a tool and start thinking about it as an operational layer.
Milan, mid-May. Thousands of people from startups, corporations and research institutions. On stage: founders of technologies used by billions, alongside CEOs rewriting entire industries as we speak. I went there to find out what I do not know. And I learned more than I expected.
The Biggest Mistake of 2025: Architecture, Not Tools
Alex Ball from Genesys opened the conference with a question that kept me up at night: “Are you scaling intelligence, or are you scaling fragmentation?” Most large companies today have dozens of AI tools, each sitting in its own silo. However, Ball says this is not a tools problem. It is an architectural problem.
For decades we built systems where each one solves one specific workflow. Then AI came along and we stuffed it into those same drawers. But AI is by nature dynamic, fluid and probabilistic. It does not need fixed boundaries between systems. It works best when those boundaries do not exist.
This resonates with what I see across my portfolio.For example, a company buys a coding copilot, a customer service chatbot, an AI tool for HR, and then wonders why it does not function as a whole. Each tool is quite smart on its own, but together they create chaos because they do not talk to each other, share no memory, and each one optimises only its own small piece.
Ball says we need an orchestration system: something that sees all agents at once, sees the human workers too, remembers the full history of every interaction, and can solve problems across the entire business, not just within one department. Then he added one number that stopped me cold: 2026 is the first year in which more than half of customer intent comes from channels companies do not own: YouTube, Reddit, ChatGPT.
Customers are not asking your helpdesk. They are asking AI. A company that ignores this becomes invisible precisely where real decisions are being made today.
Claudio Ricci from TIM Enterprise named four systematic mistakes.
First: buy licences and hope. No implementation strategy, no process change, no people development.
Second: too much ROI focus. “ROI measures the value of today’s technology, not the value of the knowledge you gain along the way.” In other words: if every experiment must first prove its return, you will never reach breakthroughs.
Third: new tools, old processes. “Like putting a Ferrari engine into a horse-drawn carriage.” You have the power, but pulling on the reins will never let you use it.
And finally: not changing the organisation. “If you quadruple someone’s AI co-workers overnight without changing processes, that person becomes the bottleneck.” The result is a paradox: the more AI you deploy badly, the more work you create.
Ricci closed with a sobering number: 20% of companies capture 75% of the value from AI. The gap is growing. And it is growing fast.
Experiment Is the New Investment
Scott Likens from PwC brought a concept I intend to put on my office wall: Return on Experiment over Return on Investment. Short cycles, one to four days. Not fail fast, but learn fast. The difference matters: fail fast accepts failure, learn fast converts it into knowledge that has value even when the project does not work.
A creative agency that had calculated the ROI on AI video in 2023 would have said no. It looked like a nightmare back then. Nevertheless, the one that tried it anyway had a two-year head start by 2025 when the technology began to work. ROI would have cost them that advantage and no retrospective investment could have bought it back.
Luca Mastella, CEO of Learnn (bootstrapped company, 280,000 users, zero external capital), added a line that stuck with me: “We all use the same models, the same tools, but we get different results. The difference is not the tools. The difference is US.” It is an uncomfortable thought, because it removes the excuse of blaming the technology. AI results depend on how well you understand the problem, how precisely you prompt, how critically you evaluate the output, and how quickly you learn from mistakes.
Mastella added numbers that speak for themselves: 95% of companies cannot measure the return on AI, 77% want to train their people, but only 13% actually do it. In short, intent without action is just a wish.
Everyone Can Build. And That Changes Everything.
Aino Bergius from Lovable (valuation 6.6 billion dollars) showed a grid: 2,500 squares, each representing 3.2 million people. Just 14 orange ones, 0.6% of the world’s population, had been deciding since the nineties what software looks like: who gets an app, what features it has, how the onboarding works. Everyone else depended on what those people created. The barrier to entry was too high: years of learning, capital, team, time. Today: a laptop, WiFi, an idea. A product in hours.
There are now 40 million projects on the Lovable platform. Sabrina from São Paulo, with no technical background, built an app for checking criminal records for employers. A sales leader at Uber Eats created a presentation generator without waiting for IT. They did not have to convince anyone, wait for approval, or ask for permission. They simply built it.
Michele Catasta from Replit added a story that moved me more than any statistic: a mother with a child with dyslexia. Every app on the market punished mistakes. Marked them in red, highlighted them, corrected them conspicuously. And the child rejected them. The mother built her own spelling app, with no programming experience, where mistakes do not look frightening. The child started writing. Catasta said: “I wish I could have done this fifteen years ago.” Today, anyone can.
For the companies in my portfolio, this has one concrete implication: your team does not have to wait for IT. Does not have to wait for budget. Does not have to wait for a project approval. They have the tools. The question is whether you will let them use them.
Speed, What Is Coming, and Why Knowing Is Not Enough
Llion Jones, co-founder of Sakana AI and co-author of “Attention is All You Need” with 250,000 citations, the paper on which ChatGPT, Claude and Gemini are built, said something that genuinely surprised me: “Stop rearranging components. You will not find the next breakthrough by reshuffling existing parts.
The next big thing will not look like a transformer at all.” The context is how the transformer came to exist: not as a brilliant idea from a whiteboard, but as a practical response to hardware constraints. Recurrent networks were slow on Google’s new chips, so Jones and his colleagues simply removed the recurrence. The result was a thousandfold speedup. The lesson is not in the architecture, but in the approach: the best breakthroughs are answers to specific constraints, not the result of combining what already exists. And his company Sakana AI is working on post-transformer architecture right now. The man who invented the transformer says the era of transformers will end.
The Operational Layer Is Coming
Likens from PwC added: within six to twelve months, world models will arrive beyond language models, humanoid robots will enter factories, and agents will become the operational layer. In other words, not a tool, but the layer through which we do everything. This is a fundamental shift: until now we have used AI as an assistant we give a task to and wait for the result. The operational layer means AI does not answer queries. It perceives continuously, decides continuously, acts continuously. And then he said the line that stayed with me through the entire conference: “The human is the loop, not the human in the loop.” In other words: AI is the means, you are the goal. You will only get superpowers if you maintain control, not if you let yourself be carried by the current of automation.
What I Am Taking Back to My Portfolio Companies
First: stop buying tools and start designing architecture. For every new AI project, ask yourself: how does this fit into the whole? Who orchestrates it? Where is the data?Therefore, an isolated tool, however good, will not add nearly as much value as a well-connected system of average tools.
Second: ROI is not the right metric for early-stage AI experiments. The right metric is the speed of learning. Run short cycles and collect knowledge, not just results. Consequently, the company that experiments and learns today will have a lead next year that money cannot buy.
Third: Giada Franceschini from Boosha AI put it best: “Automating today does not mean removing human judgement. It means putting it in the right place.” Your people should move from execution to design: from completing tasks to designing the systems that complete them. If that shift does not happen, they will become the bottleneck. Not because they are slow, but because they will be overwhelmed with decisions the system should be making.
Fourth, and perhaps most importantly: Silvia Wang, CEO of Serenis, said something that reminded me why we actually do this: “The risk is not that machines will become like us. The risk is that we will slowly become like machines.” AI is here to remove tasks, not to remove meaning. Companies that understand this will not just be more efficient. They will be better places to work and places customers genuinely want to come back to.
To sum up: Milan did not give me answers. It left me with better questions. And that is exactly what I needed.
